Hacker News: remilouf

New comment by remilouf in "The M×N problem of tool calling and open-source models"

remilouf — Wed, 15 Apr 2026 09:09:18 +0000

> Ironically LLMs solve the MxN problem he's complaining about

Enlighten me please

New comment by remilouf in "The M×N problem of tool calling and open-source models"

remilouf — Tue, 14 Apr 2026 17:05:53 +0000

Ooops sorry

New comment by remilouf in "The M×N problem of tool calling and open-source models"

remilouf — Tue, 14 Apr 2026 16:14:51 +0000

Author here. You're right, it's not a hard problem, but a particularly annoying one.

The M×N problem of tool calling and open-source models

remilouf — Thu, 09 Apr 2026 15:07:13 +0000

Article URL: https://www.thetypicalset.com/blog/grammar-parser-maintenance-contract

Comments URL: https://news.ycombinator.com/item?id=47704729

Points: 159

# Comments: 50

New comment by remilouf in "I run my company from Emacs"

remilouf — Fri, 03 Apr 2026 13:45:59 +0000

I haven't always done this, and the knowledge base used to visibly degrade over time. Reviewing a PR does not take a long time, maybe a few minutes, and this compounds over time.

I run my company from Emacs

remilouf — Fri, 03 Apr 2026 12:40:21 +0000

Article URL: https://www.thetypicalset.com/blog/emacs-as-ceo-operating-system

Comments URL: https://news.ycombinator.com/item?id=47626006

Points: 20

# Comments: 1

Every AI Integration Is Held Together with Parsing Logic and Prayer

remilouf — Thu, 10 Jul 2025 11:45:40 +0000

Article URL: https://blog.dottxt.co/do-one-thing-well.html

Comments URL: https://news.ycombinator.com/item?id=44519905

Points: 2

# Comments: 0

New comment by remilouf in ".txt raises $11.9M to make language models programmable"

remilouf — Sun, 20 Oct 2024 15:17:31 +0000

This is actually pretty funny.

New comment by remilouf in "Show HN: Infinite Testimonials with FastHTML, Outlines, and Modal"

remilouf — Tue, 10 Sep 2024 19:41:24 +0000

That’d be a pretty inefficient way to generate bullshit at scale

New comment by remilouf in "[dead]"

remilouf — Thu, 02 May 2024 14:53:48 +0000

LLM evaluations are very sensitive to the details of the prompt's structure. This post shows how using structured generation reduces the results' variance and the ranking shifts.

New comment by remilouf in "Tool Use (function calling)"

remilouf — Fri, 05 Apr 2024 17:19:32 +0000

Looks like it’s quite the opposite: http://blog.dottxt.co/performance-gsm8k.html

New comment by remilouf in "Structured Generation Improves LLM Performance: GSM8K Benchmark"

remilouf — Fri, 15 Mar 2024 14:21:54 +0000

What do you mean by "semantic dimension"?

New comment by remilouf in "Structured Generation Improves LLM Performance: GSM8K Benchmark"

remilouf — Fri, 15 Mar 2024 14:06:10 +0000

That whole structured generation line of work looks promising. I hope someone else takes this and runs evaluations on other benchmarks. Curious to see if the results translate!

New comment by remilouf in "LoRAX and Outlines: Better JSON Extraction with Structured Generation and LoRA"

remilouf — Tue, 05 Mar 2024 18:24:46 +0000

Awesome work! I am really impressed by how much structured generation improves model performance.

New comment by remilouf in "Tokenization Is a Problem for LLMs"

remilouf — Sun, 04 Feb 2024 16:13:07 +0000

This article presents a way to make structured generation with LLMs much faster than standard generation, but what I find most interesting is how it highlights the issues that tokenization entails towards the end.

Tokenization Is a Problem for LLMs

remilouf — Sun, 04 Feb 2024 16:13:07 +0000

Article URL: http://blog.dottxt.co/coalescence.html

Comments URL: https://news.ycombinator.com/item?id=39251433

Points: 2

# Comments: 1

New comment by remilouf in "Show HN: VLLM with JSON Guided Generation"

remilouf — Fri, 22 Dec 2023 16:20:19 +0000

We already support regex-guided generation in the library, and could easily make an API to serve this as well if that's a feature people want!

Show HN: Vectorize OpenAI API calls (like NumPy)

remilouf — Mon, 13 Nov 2023 16:46:05 +0000

Outlines is a Python library that focuses on text generation with large language models. While the library mainly focuses on open source models we have an integration with the OpenAI API. When trying to implement Tree of Thoughts we became frustrated with the iterative workflow in which other libraries forced us. What we wanted was to be able to pass an array of prompts or completions with arbitrary shape, ask for several samples, and get an array of completions back.

So we implemented a vectorization routine inspired by NumPy's. This makes our OpenAI integration have the "broadcast" behavior described above. All call are performed concurrently.

We thought this could be of use to some folks here who are trying to implement complex workflows with the OpenAI API.

Comments URL: https://news.ycombinator.com/item?id=38252197

Points: 3

# Comments: 0

New comment by remilouf in "Show HN: LLMs can generate valid JSON 100% of the time"

remilouf — Thu, 17 Aug 2023 21:47:39 +0000

It is currently limited by the time it takes to build the index. There are obvious optimizations we can apply to this, however in a production setting it does not matter much since you only need to build the index once for each (schema, vocabulary) pair.

New comment by remilouf in "Show HN: LLMs can generate valid JSON 100% of the time"

remilouf — Thu, 17 Aug 2023 21:44:09 +0000

You mean nested JSON? It's totally possible.